What it is
The panel sits on the right of the editor and is open by default, 300 px wide (drag the divider between 240 and 520). Its header reads “✨ AI chat”, with “working on ‘<document>’” beneath it, so you always know which document a request lands on. Ask a question and you get prose, with step references by line number counted from 1. Ask for a change and you get a proposal headed “Here’s the change — nothing is applied until you say so:” — a compact diff with +added and −removed counts, the changed lines with one line of context, and Apply and Discard buttons underneath. Long diffs fold after 14 rows and say how many more rows the document has. If the reply equals the document, the chat says “The document already does that — no changes needed.”
The agent never writes on its own. A proposal waits in the conversation until you decide, and the document you are editing stays exactly as it was while you read it. While the model works, the panel cycles “Reading the process…”, “Thinking it through…” and “Writing…” every 1.4 seconds. It works on process documents and sequence diagrams alike.
How a proposal is built
Every request carries the whole current document (up to 60,000 characters — a longer one is refused, never cut), the recent conversation, summaries of older stretches and the names of your other documents, so a .do or .go the agent writes can only point at something real. The model replies in one of three modes: EDIT returns the complete revised document, NEWDOC returns a separate new document with a “NAME:” line, ANSWER returns prose. Runic diffs an EDIT against your text on the client, so what you review is the real change, not a description of it. An EDIT is told to copy every line it was not asked to touch verbatim — text, captions, notes, indentation and order.
Before you see an EDIT or a NEWDOC, the reply goes through the same parser and structural audit the editor uses, for processes and for sequences. Only findings the model introduced count: your own document is the baseline, so a one-branch decision you wrote on purpose is not “fixed” behind your back. If the reply scores worse than your text — a syntax error, a decision with one branch, a .back that points nowhere, an .end that closes the wrong block — the model gets one correction pass with the validator report, and the fix is kept only when it scores better, adds no parse errors, was not cut short and still contains every original line. When the correction was kept, the reply carries a small “self-checked” tag under the message, next to the model label.
It asks before it assumes
An ambiguous change request gets a question, not a guess. If it is unclear which of several similar steps you mean, which branch, or what threshold, role, system, SLA or retry count you have in mind — or you name something that is in neither the document nor your workspace — the agent answers with one precise question that lists the candidates, quoting the step texts. Cosmetic ambiguity it resolves on its own. A value you did not give is never invented: the edit carries a “.note TBD: <what is missing>” line, in the document’s language, where the number or name should go.
The same rule covers memory. “Do that” or “apply those” builds the edit from the suggestions earlier in the conversation and applies all of them. But when the detail it needs lives only in an archived summary that was not recalled, or in a history message that ends with “[… truncated]”, the agent says which detail it needs instead of filling the gap.
One tier per job
Ordinary turns run on the mid tier: Claude Sonnet 5 with thinking switched off and a 3,000-token reply budget. A request is routed to the heavy tier when it is whole-document work: it carries an attachment; it contains a whole-flow word such as redesign, rewrite, restructure, from scratch, overhaul, audit, architecture, end-to-end, “design a” or “create a” (in English or Turkish — baştan, sıfırdan, yeniden tasarla, tüm akış, denetle and more); it is longer than 600 characters; it asks to fix, improve or clean a document over 80 lines; or it asks to review or analyse (review, analyse, incele, denetle, gözden geçir) a document over 60 lines. The heavy tier is Sonnet 5 too, but with adaptive thinking on, medium effort and a 4,000-token budget plus headroom for the thinking. Opus 5 only runs when the owner sets ANTHROPIC_MODEL_HEAVY; on your own key, the “Main model” you named answers both tiers. The label under each reply — SONNET, HAIKU or OPUS for Claude, the model id otherwise — tells you which model answered.
Questions that need outside knowledge — regulations, GDPR or KVKK, ISO standards, industry best practice, benchmarks, “how do banks do this” — switch on web search for that call, capped at three searches. An answer that used it is tagged “web research” and ends with a “Sources:” line. Document bodies never contain links or citations, and the built-in deep review never searches.
Apply is guarded
Apply only works while the proposal’s document is open and still equals the text the proposal was computed against. Edit anything in between, or switch documents, and the button greys out with the note “document changed”; undo the edit and it lights up again. A proposal that is still pending when its messages are archived, or whose document has been deleted, is marked “Superseded — the document moved on” instead of being forced in. On a team document where you are a viewer, the agent tells you to ask an editor.
When you do apply, the change types itself in from the first line that differs, so the diagram redraws step by step rather than snapping. The entry flips to “Applied ✓” and gains a ↩ Restore button — the checkpoint. On a process, the automatic structural check then posts anything that is still off.
Guards and limits
The chat needs a signed-in account with a verified email and allows 60 messages an hour per account (“Chat limit reached — try again in an hour.”). A message is read up to 4,000 characters. A document over 60,000 characters is refused with “This document is too long for chat (N of 60,000 characters) — split it into linked processes.” The reply budget follows the document’s size, so a long flow is never cut mid-branch; a reply that still hits its ceiling is retried once with double the budget (up to 24,000 tokens) when at least 60 seconds of the route’s 150-second limit remain, and a second cut-off comes back as “The reply was too long to finish in one go — select the part to change, or split the document into linked processes.”
A shrink guard protects your steps: when an EDIT returns fewer than half the non-empty lines of a document of four lines or more, and your message contained no shortening verb (remove, delete, trim, merge, simplify, rewrite, sil, kısalt, birleştir and the like — a negation such as “don’t remove” cancels it), no proposal is made and the chat says “The AI returned N lines for your M-line document — it may have dropped steps, so no change was proposed. Try again, or select the lines to change and use AI edit.” Errors post as “⚠ <message>”. The audit button, the diagram preview and the selection chip exist for processes only; the self-check runs for both kinds. Fullscreen on the diagram hides the panel along with the editor; Esc brings them back.
How to use it
- Open the panelIt is open by default on the right. If it is collapsed, click the ✨ rail on the right edge (tooltip “Open AI chat”) or “AI chat” in the top bar; a dot on the rail means the conversation has messages.
- Say what you wantType in the box labelled “Ask, or describe a change…” and press Enter. Shift+Enter adds a line; the Send button does the same as Enter and is inactive while the agent is busy.
- Read the replyA question gets prose. A change gets a diff with +/− counts. An ambiguous request gets one question listing the candidates — answer it and the edit follows. The model label and any “web research” or “self-checked” tags sit under the message.
- Apply or DiscardClick Apply to write the change in, or Discard to drop it. While a process proposal is pending the entry says “previewing in the diagram →” and the flow pane shows it.
- Come back laterEvery applied entry keeps a ↩ Restore button that returns the document to how it was before that change.
- Start fresh when you want“Clear” in the panel header wipes the conversation, including archived stretches and their checkpoints.
Commands & shortcuts
| Enter | send the message |
| Shift+Enter | new line inside the message |
| Apply / Discard | accept or drop a pending proposal |
| ↩ Restore | return to the state before an applied change |
| Clear | wipe the conversation and its checkpoints |
| ✨ rail · AI chat | reopen the collapsed panel; the collapse icon in the panel header folds it away |
| .note TBD: … | what the agent writes where you gave no value, instead of inventing one |
| SONNET · web research · self-checked | the meta line under a reply: which model, whether it searched, whether the repair pass changed the text |
Tips
- Say “do what you suggested” — the agent builds the edit from the conversation instead of asking you to repeat it.
- Use the words “redesign” or “from scratch” when you want a whole-document rewrite; that routes the turn to the heavy tier, Sonnet 5 with thinking on.
- Give the number when you have it: “above 5,000 the CFO approves” lands as an .if, while “above some threshold” lands as a .note TBD you will have to fill in.
- Ask “what does GDPR require here?” style questions to get a researched answer with a Sources line.
- For a change to a few specific lines, select them and use Copy to chat — the edit is then confined to that region, and a long document stops being a problem.
- If Apply is greyed out with “document changed”, ask again; the new proposal is computed against the current text.
Questions
Does the agent ever change my document without asking?
No. Change requests come back as proposals; only Apply writes. The automatic structural check after an applied change posts findings, it never edits.
Which model answers?
On Runic’s key, Claude Sonnet 5 in both cases: with thinking off for ordinary turns, with adaptive thinking on for whole-document work. Opus 5 appears only when the owner has set ANTHROPIC_MODEL_HEAVY. On your own key it is the Main model you named. The label under each reply says which one ran.
Why did it ask me a question instead of making the change?
Because the request could be read more than one way in a way that changes the diagram — several steps matched, a branch or a threshold was unstated, or a name matched nothing in your workspace. Answer the question and the edit follows; the agent will not guess.
Does the chat work on sequence diagrams?
Yes: questions, diff proposals and the self-check work on the open sequence diagram. The Audit button, the diagram preview and the selection chip only exist for processes.
What happens if I edit while a proposal is pending?
Apply greys out with “document changed” and stays that way until the document equals the proposal’s base text again. It never overwrites your newer edits. Ask again to get a fresh proposal.